ETL性能

转自：http://en.wikipedia.org/wiki/Extract,_transform,_load

Performance

ETL vendors benchmark their record-systems at multiple TB (terabytes) per hour (or ~1 GB per second) using powerful servers with multiple CPUs, multiple hard drives, multiple gigabit-network connections, and lots of memory. The fastest ETL record is currently held by Syncsort,^[1]Vertica and HP at 5.4TB in under an hour which is more than twice as fast as the earlier record held by Microsoft and Unisys.

In real life, the slowest part of an ETL process usually occurs in the database load phase. Databases may perform slowly because they have to take care of concurrency, integrity maintenance, and indices. Thus, for better performance, it may make sense to employ:

• Direct Path Extract method or bulk unload whenever is possible (instead of querying the database) to reduce the load on source system while getting high speed extract
• most of the transformation processing outside of the database
• bulk load operations whenever possible.

Still, even using bulk operations, database access is usually the bottleneck in the ETL process. Some common methods used to increase performance are:

• Partition tables (and indices). Try to keep partitions similar in size (watch for null values which can skew the partitioning).
• Do all validation in the ETL layer before the load. Disable integrity checking (disable constraint ...) in the target database tables during the load.
• Disable triggers (disable trigger ...) in the target database tables during the load. Simulate their effect as a separate step.
• Generate IDs in the ETL layer (not in the database).
• Drop the indices (on a table or partition) before the load - and recreate them after the load (SQL: drop index ...; create index ...).
• Use parallel bulk load when possible — works well when the table is partitioned or there are no indices. Note: attempt to do parallel loads into the same table (partition) usually causes locks — if not on the data rows, then on indices.
• If a requirement exists to do insertions, updates, or deletions, find out which rows should be processed in which way in the ETL layer, and then process these three operations in the database separately. You often can do bulk load for inserts, but updates and deletes commonly go through an API (using SQL).

Whether to do certain operations in the database or outside may involve a trade-off. For example, removing duplicates using distinct may be slow in the database; thus, it makes sense to do it outside. On the other side, if using distinct will significantly (x100) decrease the number of rows to be extracted, then it makes sense to remove duplications as early as possible in the database before unloading data.

A common source of problems in ETL is a big number of dependencies among ETL jobs. For example, job "B" cannot start while job "A" is not finished. You can usually achieve better performance by visualizing all processes on a graph, and trying to reduce the graph making maximum use of parallelism, and making "chains" of consecutive processing as short as possible. Again, partitioning of big tables and of their indices can really help.

Another common issue occurs when the data is spread between several databases, and processing is done in those databases sequentially. Sometimes database replication may be involved as a method of copying data between databases - and this can significantly slow down the whole process. The common solution is to reduce the processing graph to only three layers:

• Sources
• Central ETL layer
• Targets

This allows processing to take maximum advantage of parallel processing. For example, if you need to load data into two databases, you can run the loads in parallel (instead of loading into 1st - and then replicating into the 2nd).

Of course, sometimes processing must take place sequentially. For example, you usually need to get dimensional (reference) data before you can get and validate the rows for main "fact" tables.

[edit]Parallel processing

A recent development in ETL software is the implementation of parallel processing. This has enabled a number of methods to improve overall performance of ETL processes when dealing with large volumes of data.

ETL applications implement three main types of parallelism:

• Data: By splitting a single sequential file into smaller data files to provide parallel access.
• Pipeline: Allowing the simultaneous running of several components on the same data stream. For example: looking up a value on record 1 at the same time as adding two fields on record 2.
• Component: The simultaneous running of multiple processes on different data streams in the same job, for example, sorting one input file while removing duplicates on another file.

All three types of parallelism usually operate combined in a single job.

An additional difficulty comes with making sure that the data being uploaded is relatively consistent. Because multiple source databases may have different update cycles (some may be updated every few minutes, while others may take days or weeks), an ETL system may be required to hold back certain data until all sources are synchronized. Likewise, where a warehouse may have to be reconciled to the contents in a source system or with the general ledger, establishing synchronization and reconciliation points becomes necessary.